JVET-AD0241 AHG13: On Film Grain Objective Metrics [X. Meng, W. Zhang, S. Labrozzi (Disney Streaming)]
This proposal aims to explore the feasibility of NSS (Natural Scene Statistics) as an objective metric in the film grain related studies in JVET. Specifically, the capability of NSS on measuring the quality and fidelity of video with film grain was evaluated. Given the fact that there is no widely adopted objective metric specific to film grain, the investigation in this proposal was designed to address the following questions,
- Can the characteristics of film grain be abstracted and represented by NSS in a quantitive way?
- How are the NSS results related to the subjective perception of human eyes?
- Does the NSS work universally, or with limitations and/or conditions?
According to the experimental results, NSS based method showed promising outcomes to measure the film grain, especially when the content is relatively simple, and the film grain characteristics are consistent. Reliable results correlated to subjective perception were obtained.
However, it’s also observed that NSS is prone to be disturbed if the content is with high spatial complex regions. Besides, it’s also sensitive to the pooling method when applying NSS to video. For example, averaged NSS among multiple frames didn’t work well in the tests.
Comparisons were performed when modifying parameters such as strength and cutoff frequency of the synthesis, but in some cases no differences were identified by the metric, even though visually different.
Experiments mostly done with small cropped regions.
Current input of the metric is pictures with film grain. Could the similarity check potentially be improved when content and film grain would be available separately, like in the ground truth approach? This would need to be investigated.
It would be interesting to investigate how reliable the metric is in its assessment, which could however only be assessed with more material.
How was precisely the modification of the SEI message parameters performed? Were the parameters originally extracted appropriately reconstructing the film grain? This should be taken into account when concluding about the benefit of the metric.
More study and future contributions were welcome.